Big Self-Supervised Models Advance Medical Image Classification (2101.05224v2)

Abstract: Self-supervised pretraining followed by supervised fine-tuning has seen success in image recognition, especially when labeled examples are scarce, but has received limited attention in medical image analysis. This paper studies the effectiveness of self-supervised learning as a pretraining strategy for medical image classification. We conduct experiments on two distinct tasks: dermatology skin condition classification from digital camera images and multi-label chest X-ray classification, and demonstrate that self-supervised learning on ImageNet, followed by additional self-supervised learning on unlabeled domain-specific medical images significantly improves the accuracy of medical image classifiers. We introduce a novel Multi-Instance Contrastive Learning (MICLe) method that uses multiple images of the underlying pathology per patient case, when available, to construct more informative positive pairs for self-supervised learning. Combining our contributions, we achieve an improvement of 6.7% in top-1 accuracy and an improvement of 1.1% in mean AUC on dermatology and chest X-ray classification respectively, outperforming strong supervised baselines pretrained on ImageNet. In addition, we show that big self-supervised models are robust to distribution shift and can learn efficiently with a small number of labeled medical images.

Overview of "Big Self-Supervised Models Advance Medical Image Classifications"

The paper "Big Self-Supervised Models Advance Medical Image Classifications" presents a comprehensive paper on leveraging self-supervised learning (SSL) for medical image classification tasks. It addresses the critical issue of scarce labeled data in medical image analysis by employing SSL methods that utilize large amounts of unlabeled data for pretraining.

Key Contributions

1. Self-Supervised Pretraining: The paper demonstrates the effectiveness of self-supervised learning techniques for medical image classification. It highlights that SSL on datasets like ImageNet, followed by additional pretraining on domain-specific medical images, enhances the accuracy of classifiers significantly when compared to conventional supervised pretraining strategies.
2. Multi-Instance Contrastive Learning (MICLe): Introducing MICLe, the paper extends existing contrastive learning models by utilizing multiple images of the same pathology to create more informative positive pairs. This approach capitalizes on multi-instance data, which is commonplace in medical image datasets, to learn robust representations.
3. Empirical Results: The models pretrained with SSL showed a notable improvement of 6.7% in top-1 accuracy on dermatology classification and a 1.1% increase in mean AUC for chest X-ray classification over strong supervised baselines.
4. Robustness to Distribution Shift: The paper also highlights that SSL models maintain robust performance against distribution shifts, making them reliable for real-world clinical applications.

Methodology

The research employs a sequential approach:

• Initial self-supervised pretraining on ImageNet using SimCLR.
• Additional pretraining on domain-specific datasets, utilizing the proposed MICLe method.
• Supervised fine-tuning on labeled medical images tailored to specific tasks.

This methodology effectively bridges the domain gap and scales well without needing labeled data during the initial stages.

Theoretical and Practical Implications

• Theoretical: The paper underscores the efficacy of contrastive learning and its adaptability to multi-instance data, thus broadening the theoretical understanding of self-supervised strategies in complex domains like medical imaging.
• Practical: The implications for medical imaging are significant. The ability to train models with limited labeled data while still achieving high accuracy can lead to broader adoption of AI in healthcare, potentially alleviating the need for extensive manual annotations.

Future Directions

The research opens several avenues for future exploration:

• Scaling SSL methods with larger unlabeled datasets could further enhance performance.
• Investigating SSL transferability across different medical imaging modalities could provide insights into developing more generalizable models.
• Combining SSL with semi-supervised learning strategies may optimize label efficiency even further.

Conclusion

The paper makes substantial contributions to the field of medical image analysis by demonstrating the potential of self-supervised learning as a robust pretraining strategy. Its innovative use of MICLe and impressive empirical results suggest that SSL can significantly outperform traditional supervised methods, paving the way for more efficient and effective AI applications in medical diagnostics.